The Economics of Intelligence

Several announcements over the past two weeks point in the same direction:

• Moonshot released Kimi K3, the first open-weight model to reach 2.8 trillion parameters, performing at near-frontier levels against Claude Opus 4.8 and GPT-5.5. 

• GLM 5.2 pushed the price-performance curve further, offering competitive capability at a fraction of frontier pricing. 

• Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling… a 975 billion parameter open-weight model designed for customization and fine-tuning rather than benchmark dominance. Open weights. Apache 2.0 license. 

• Microsoft announced the RTX Spark, a laptop purpose-built to run capable models locally, and GitHub Copilot now routes between frontier, open-weight, and local models automatically. Satya Nadella’s framing: “unmetered intelligence to every desk and every home.” 

Taken together, they tell the same story from different layers of the stack: models, pricing, hardware, and orchestration.

We’re already seeing this playing out. On OpenRouter, a major marketplace where developers access hundreds of competing AI systems, Chinese models regularly occupy the top spots by weekly token usage, with GLM 5.2 already in the top five. While OpenRouter isn’t a proxy for the entire enterprise market, it is one of the clearest real-time signals of developer adoption. That’s revealed preference. Developers are voting with actual workloads.

Ben Thompson framed the economics this week in a way that clarifies why. Tokens, he argues, are not the right unit of competition. A model that produces the correct answer using half the tokens is effectively cheaper, even if its per-token price is higher. What ultimately matters is the outcome. Once you measure by outcome intelligence rather than tokens, the question becomes who can produce that intelligence at the lowest cost.

The pattern is emerging from every layer of the stack. Open-weight models approaching frontier capability are here. Inference pricing continues to fall. Local execution is becoming viable for enterprise workflows. Model routing is automating the selection of the right capability at the right cost for each task.

With these, Jevons Paradox playing out faster than many expected. When the cost of a critical input declines dramatically, consumption tends to increase rather than decrease because entirely new applications become economically viable. Cheaper intelligence doesn’t simply reduce spend. It expands the set of problems worth solving.

For media, this changes the economics and approach. Content analysis, metadata enrichment, QC automation, compliance checking, rights validation, localization, and packaging can run continuously against an entire library, with costs that decline quarter over quarter rather than accumulate linearly with every operation. Open-weight models fine-tuned to specific domains, taxonomies, compliance rules, and editorial standards become a persistent part of an operations stack rather than an expensive per-transaction service.

As frontier capability becomes broadly available, differentiation moves elsewhere... the orchestration, the domain-specific context, the proprietary data, the workflow architecture, and the feedback loops that make intelligence useful for specific operations. The scarce resources become proprietary context, operational execution, and organizational design. Companies that begin architecting around that today are likely to build structural advantages over those still budgeting as if intelligence is a scarce, expensive resource to be rationed.